I spent two hours with my AI agent building a product launch plan. We mapped out the steps, assigned tasks, set timelines. Then the context window filled up, the system compacted the conversation, and the plan disappeared. The agent said "got it." Nothing changed.
This happened maybe forty times before I understood the problem.
The Problem
AI agents don't have memory. They have the illusion of memory. They reference earlier conversation, they build on what you said, they act like they know what's going on. Then the context resets and all of that was resting on nothing.
The instinct is to write better prompts. More instructions. "Always remember to check the plan." This doesn't work. You're taping notes to the forehead of someone with amnesia.
The Fix
Stop relying on the agent to remember. Write everything to files on disk, automatically, before it can be forgotten.
I built a system called Project Pulse. Three layers.
Layer 1: Extraction Cron. Every two hours, a cron job reads the last 30 messages from the conversation. For each project discussed, it answers 7 questions:
- What was completed?
- What are the next actions?
- What's blocked?
- What new ideas came up?
- Were any recurring processes established?
- What numbers should we track?
- Did I correct something or express a preference?
The answers get edited into that project's PROJECT.md file. Not appended to a log. Edited into the specific section, with a date stamp. The file always reflects the current state of the project. When the agent starts a new session, it reads the file. It doesn't need to remember. It's written down.
Layer 2: Pre-Compaction Flush. When the context window hits ~70% capacity, the agent automatically runs the same extraction before the system compacts. Safety net. Even if the cron hasn't fired, the flush catches everything before it disappears.
Layer 3: "Save This." When I say "save this" mid-conversation, the agent immediately extracts and writes to the relevant files. Three words. Instant.
The Structure
Every project gets a PROJECT.md with standardized sections: Progress, Active Tasks, Blockers, Ideas, Recurring Ops, Monitor, Preferences. The 7 extraction questions map directly to these sections.
Then there's PROJECT-PULSE.md. A dashboard. One line per project with status, next actions, and blockers. If a project has been yellow for five days, it escalates. Projects don't die silently.
What Changed
Before this, I'd explain a plan to my agent, leave, come back, and she'd have no idea what I was talking about. Now the plan is on disk. The context can compact all it wants. The files persist.
I packaged this as an OpenClaw skill called project-pulse. Install it and you get the extraction prompt, the cron setup, the PROJECT.md template, and the dashboard. The 7 questions are the core of it. Specific enough to capture what matters, structured enough to apply consistently.
If you're building anything serious with an AI agent, set up memory before anything else. None of the fancy tools matter if your agent forgets why it's doing what it's doing.